Macenko PCA

High-performance stain matrix estimation and colour deconvolution for histology images using the Macenko PCA method, with the compute-intensive core written in Rust via PyO3.

Supported platforms: Linux and macOS only. Windows is not supported due to differences in BLAS/LAPACK / OpenBLAS builds that can cause numerical and linking inconsistencies; to ensure reproducible numerical results we build and test only on Linux and macOS.

This implements the method described in:

Macenko, M. et al. "A method for normalizing histology slides for quantitative analysis." ISBI 2009.

And was based off the implementation from HistomicsTK

https://github.com/DigitalSlideArchive/HistomicsTK

Performance improvements between the HistomicsTK implementation and this one are discussed here. We saw almost a quarter of the RAM usage and up to a 7x speedup in real world examples.

Features

Quick Start

Installation

pip install macenko-pca

Full Workflow Example

import numpy as np
from macenko_pca import (
    rgb_separate_stains_macenko_pca,
    rgb_color_deconvolution,
    reconstruct_rgb,
)

# Load or create an RGB image (H×W×3, values in [0, 255])
im_rgb = np.random.rand(256, 256, 3) * 255.0

# 1. Estimate the 3×3 stain matrix from the image
stain_matrix = rgb_separate_stains_macenko_pca(im_rgb)
print("Stain matrix:\n", stain_matrix)

# 2. Decompose the image into per-stain concentration channels
concentrations = rgb_color_deconvolution(im_rgb, stain_matrix)

hematoxylin = concentrations[:, :, 0]
eosin = concentrations[:, :, 1]
residual = concentrations[:, :, 2]

# 3. Modify concentrations (e.g. isolate hematoxylin only)
concentrations_h_only = concentrations.copy()
concentrations_h_only[:, :, 1] = 0.0  # zero-out eosin
concentrations_h_only[:, :, 2] = 0.0  # zero-out residual

# 4. Reconstruct back to RGB
im_hematoxylin_only = reconstruct_rgb(concentrations_h_only, stain_matrix)

Step-by-Step API

from macenko_pca import (
    rgb_to_sda,
    separate_stains_macenko_pca,
    color_deconvolution,
    reconstruct_rgb,
)

# Convert RGB to SDA (Stain Density Absorbance) space
im_sda = rgb_to_sda(im_rgb)

# Estimate stain vectors from the SDA image
stain_matrix = separate_stains_macenko_pca(im_sda)

# Decompose SDA image into stain concentrations
concentrations = color_deconvolution(im_sda, stain_matrix)

# Reconstruct RGB from (possibly modified) concentrations
im_reconstructed = reconstruct_rgb(concentrations, stain_matrix)

Half the RAM — Use float32

# Simply cast your input to float32 — the Rust backend will use f32 throughout
im_rgb_f32 = im_rgb.astype(np.float32)
stain_matrix = rgb_separate_stains_macenko_pca(im_rgb_f32)       # f32
concentrations = rgb_color_deconvolution(im_rgb_f32, stain_matrix)  # f32
reconstructed = reconstruct_rgb(concentrations, stain_matrix)        # f32

Precision / dtype rules

The dtype of your input array controls which Rust code path is taken:

The return array's dtype always matches the computation dtype.

API Reference

Stain Matrix Estimation

rgb_separate_stains_macenko_pca(im_rgb, ...)

End-to-end: takes an RGB image (H, W, 3) and returns a (3, 3) stain matrix.

separate_stains_macenko_pca(im_sda, ...)

Lower-level: operates on an image already in SDA space.

Colour Conversion

rgb_to_sda(im_rgb, ...)

Converts an RGB image or matrix to SDA (stain-density-absorbance) space.

Colour Deconvolution (Applying Stain Vectors)

color_deconvolution(im_sda, stain_matrix)

Decomposes an SDA image into per-stain concentration channels using the inverse of the stain matrix. Each output channel i holds the concentration of stain i.

rgb_color_deconvolution(im_rgb, stain_matrix, bg_int=None)

Convenience wrapper that converts RGB → SDA → concentrations in a single call.

Reconstruction

reconstruct_rgb(concentrations, stain_matrix, bg_int=None)

Reconstructs an RGB image from stain concentrations and a stain matrix. Inverts the deconvolution: SDA = concentrations × Wᵀ, then converts SDA back to RGB. Useful for stain normalisation workflows where you modify concentration channels and then reconstruct.

See the full API documentation for parameter details.

Development Setup

Prerequisites: Python 3.9+, a Rust toolchain, and Hatch.

On Linux you also need OpenBLAS development headers (libopenblas-dev on Debian/Ubuntu). On macOS: brew install openblas pkg-config.

git clone https://github.com/LavLabInfrastructure/macenko-pca.git
cd macenko-pca
pip install maturin hatch
maturin develop --release

Optionally install pre-commit hooks:

pip install pre-commit
pre-commit install

Common Commands

Run these directly or use the provided Makefile shortcuts (e.g. make test, make lint).

Docker

# Run tests via Docker
docker build --target maturin -t macenko-pca:maturin .
docker run --rm -e HATCH_ENV=test macenko-pca:maturin cov

# Production image (just the installed wheel)
docker build --target prod -t macenko-pca:prod .

Publishing to PyPI

This project uses trusted publishing (OIDC) — no API tokens or secrets are needed. The publish.yml workflow handles everything automatically.

One-Time Setup (PyPI)

1. Go to https://pypi.org/manage/account/publishing/ (logged in as a maintainer of the lavlab org).
2. Click "Add a new pending publisher" and fill in:
   • PyPI project name: macenko-pca
   • Owner: lavlab (the GitHub organisation)
   • Repository: macenko-pca
   • Workflow name: publish.yml
   • Environment name: pypi
3. (Optional) Repeat on TestPyPI with environment name testpypi to enable dry-run publishes.

One-Time Setup (GitHub)

1. In the repository settings, go to Environments.
2. Create an environment called pypi.
   • Optionally add a protection rule requiring manual approval before publishing.
3. (Optional) Create an environment called testpypi for test publishes.

How to Release

# 1. Bump the version in src/macenko_pca/__about__.py and Cargo.toml
# 2. Commit and tag
git add -A
git commit -m "release: v0.2.0"
git tag v0.2.0
git push && git push --tags

# 3. Create a GitHub Release from the tag (via the web UI or `gh` CLI)
gh release create v0.2.0 --generate-notes

Creating the release triggers publish.yml, which:

1. Builds wheels on Linux (manylinux) and macOS (x86_64 + arm64). Windows builds are intentionally omitted due to platform LAPACK/BLAS inconsistencies that affect numerical reproducibility.
2. Builds a source distribution.
3. Publishes everything to PyPI via trusted publishing.

Testing a Publish (Without a Release)

You can manually trigger the workflow against TestPyPI:

1. Go to Actions → Publish to PyPI → Run workflow.
2. Select testpypi as the target.
3. Verify the package at https://test.pypi.org/project/macenko-pca/.

Project Structure

macenko-pca/
├── src/
│   └── macenko_pca/            # Python package source
│       ├── __init__.py          # Public API & version export
│       ├── __about__.py         # Version string
│       ├── deconvolution.py     # Pythonic wrappers with dtype dispatch
│       └── py.typed             # PEP 561 marker
├── rust/                        # Rust source (compiled via PyO3 + maturin)
│   ├── lib.rs                   # PyO3 module entry point (f32 + f64 variants)
│   ├── float_trait.rs           # MacenkoFloat supertrait (f32/f64)
│   ├── color_conversion.rs      # RGB → SDA transform (generic)
│   ├── color_deconvolution.rs   # SDA → concentrations, RGB reconstruction
│   ├── complement_stain_matrix.rs
│   ├── linalg.rs                # SVD, magnitude, normalisation (generic)
│   ├── rgb_separate_stains_macenko_pca.rs
│   ├── separate_stains_macenko_pca.rs
│   └── utils.rs                 # Image ↔ matrix helpers
├── tests/
│   ├── conftest.py              # Shared pytest fixtures
│   └── test_deconvolution.py    # Library function tests (104 tests)
├── docs/                        # MkDocs source files
├── .github/
│   ├── workflows/
│   │   ├── build.yml            # CI: build wheels on every push/PR
│   │   ├── pytest.yml           # CI: run tests across Python versions
│   │   ├── lint.yml             # CI: ruff lint + format check
│   │   └── publish.yml          # CD: publish to PyPI on release
│   └── dependabot.yml           # Auto-update deps, actions, Docker, Cargo
├── Cargo.toml                   # Rust crate configuration
├── pyproject.toml               # Python project & tool config (maturin backend)
├── Dockerfile                   # Multi-stage build (maturin / dev / prod)
├── Makefile                     # Dev shortcuts
├── mkdocs.yml                   # Docs config
├── .pre-commit-config.yaml      # Pre-commit hooks
├── .editorconfig                # Editor consistency
└── .gitignore

Design Philosophy

This project follows a library-first approach:

1. All logic lives in importable modules under src/macenko_pca/.
2. Heavy computation is delegated to Rust (rust/) for maximum throughput — SVD via ndarray-linalg, parallelism via rayon. All Rust functions are generic over f32/f64 via the MacenkoFloat trait.
3. The Python layer (deconvolution.py) detects the input array's dtype and dispatches to the appropriate typed Rust function, providing input validation and rich docstrings.
4. Tests call library functions directly.

This keeps your code reusable whether it's called from another package, a Jupyter notebook, or a web API.

Contributing

See CONTRIBUTING.md for development guidelines.

License

macenko-pca is distributed under the terms of the MIT license.